Re: [RFC PATCH 07/11] IB/Verbs: Use management helper has_mcast() and, cap_mcast() for mcast-check

[Yun Wang <yun.wang@profitbricks.com>]

On Fri, Mar 27, 2015 at 6:05 PM, ira.weiny <ira.weiny@intel.com> wrote:
> On Fri, Mar 27, 2015 at 10:28:20AM -0600, Jason Gunthorpe wrote:
>> On Fri, Mar 27, 2015 at 04:46:57PM +0100, Michael Wang wrote:
>> >
[snip]
>> > -            if (rdma_transport_is_ib(id_priv->cma_dev->device)) {
>> > +            if (has_mcast(id_priv->cma_dev->device)) {
>>
>> This might make more sense as cap_ib_multicast / cap_ip_multicast
>
> Agreed.
>

Will be changed in next version :-)

>>
>> >                  switch (rdma_port_get_link_layer(id->device, id->port_num)) {
>> >                  case IB_LINK_LAYER_INFINIBAND:
>> >                      ib_sa_free_multicast(mc->multicast.ib);
>>
>> > diff --git a/drivers/infiniband/core/multicast.c b/drivers/infiniband/core/multicast.c
>> > index 17573ff..ffeaf27 100644
>> > +++ b/drivers/infiniband/core/multicast.c
>> > @@ -780,7 +780,7 @@ static void mcast_event_handler(struct ib_event_handler *handler,
>> >      int index;
>> >
>> >      dev = container_of(handler, struct mcast_device, event_handler);
>> > -    if (!rdma_port_ll_is_ib(dev->device, event->element.port_num))
>> > +    if (!cap_mcast(dev->device, event->element.port_num))
>> >          return;
>>
>> These should probably be cap_ib_sa - that is what they are guarding
>> against.
>>
>> But it seems redudent, since mcast_add_one will already not add a port that is
>> not IB, so mcast_event_handler is not callable. Something to do with
>> rocee/ib switching?
>
> I'm not sure about this either.  This check seems to be necessary only on a
> per-port level.  It does seem apparent that one can't go from Eth to IB.  What
> happens if you go from IB to Eth on the port?

I also feel it's redundant at first glance, but just not sure if it
could be removed, lack of some knowledge :-P

>
>>
>> >      index = event->element.port_num - dev->start_port;
>> > @@ -807,7 +807,7 @@ static void mcast_add_one(struct ib_device *device)
>> >      int i;
>> >      int count = 0;
>> >
>> > -    if (!rdma_transport_is_ib(device))
>> > +    if (!has_mcast(device))
>> >          return;
>>
>> Again, this seems redundant, every port is tested directly below, why
>> is this check needed?
>
> Agreed.  Same as my comments about the SA support.  This is really only
> needed on ports which need to register with the SA (or perhaps some future
> entity) for Mcast support.

I will recheck all the logical around has_XX() see if we can get rid of them ;-)

>
> Also this is part of the ib_sa module and exports the function
> ib_sa_join_multicast.  So that this point it is covered under the
> cap_sa(device, port) call.
>
> So the implementation of cap_mcast at this point is:
>
> cap_mcast(device, port)
> {
>         return cap_sa(device,port);
> }
>

Sounds good :-) will be in next version.

>>
>> Looking at this, I do wonder how a port can dynamically change between
>> rocee and IB.. If the link value changes then mcast_remove_one will
>> not be a perfect reversal of mcast_add_one. Bug?
>>
>> It feels necessary to understand what happens when a port dynamically
>> switches to ethernet on mlx hardware to validate these patches :(
>
> Agreed.

Maybe we can temporarily reserve the old logical, and gradually solve
these problems?

Regards,
Michael Wang


>
> :-(
>
> -- Ira
>
>>
>> Jason